This is what “thirty days” actually looks like.
WITHOUT A SPRINT
- Week 1 — AI strategy kickoff. Everyone is excited.
- Week 2 — Vendor demos. Nobody has touched real data.
- Week 4 — A chatbot over the PDFs. Nobody uses it.
- Week 7 — Security review of a tool nobody chose.
- Week 11 — Pilot plan, third revision.
- Week 16 — IT says it can't touch the ERP.
- Week 24 — New vendor. Start again.
Six months in. Still nobody has clicked anything.
WITH A SPRINT
- Day 1 — Trace the real workflow — every email, sheet and exception.
- Day 3 — One use case, its boundaries, and what it's worth.
- Day 5 — The approval screen, in code, on a live link.
- Weeks 2–4 — Agent, evals, and a shadow run on your systems.
- Shipped in week 4.
Everything past here is iteration on software real people are already using.
The thirty days, unpacked
Audit. Build. Evals. Deploy.
The timeline above is the why. This is the what — each stage earns the right to the next.
- Week 1 · Design sprint
Audit the workflow that's actually running.
Interviews, emails, spreadsheets, SOPs and every exception path, traced into one operating map: how the work happens today, how it should run with AI in it, the one use case worth doing first, and what it's worth in hours, cost and errors.
- Week 2
Build it on top of what you already run.
Rules go in code, judgment goes to an agent, and one clear decision goes to a person. Every prompt, tool call and approval is logged. It lives in your stack and your repo, not a vendor's.
- Week 3
Turn "it works" into evidence.
Evaluation cases for the normal path, the edge case, missing data, ambiguous requests and high-risk actions. You get a pass rate, the failure categories, and the rules for when it must escalate to a human.
- Week 4
Deploy in shadow, then earn autonomy.
It runs alongside your team inside your own infrastructure, with logs, alerts and rollback. A go/no-go readout for engineering and leadership, then a 14-day support window.